SPRUI04 User guide

SPRUI04F july 2015 – april 2023

5.5.2 A Dot Product Example That Avoids Memory Bank Conflicts

The C code in Example5-7 implements a dot product function. The inner loop is unrolled once to take advantage of the C6000's ability to operate on two 16-bit data items in a single 32-bit register. LDW instructions are used to load two consecutive short values. The linear assembly instructions in Example5-8 implement the dotp loop kernel. Example5-9 shows the loop kernel determined by the assembly optimizer.

For this loop kernel, there are two restrictions associated with the arrays a[ ] and b[ ]:

Because LDW is being used, the arrays must be aligned to start on word boundaries.
To avoid a memory bank conflict, one array must start in bank 0 and the other array in bank 2. If they start in the same bank, then a memory bank conflict occurs every cycle and the loop computes a result every two cycles instead of every cycle, due to a memory bank stall. For example:
Bank conflict:
```
      MVK   0, A0
   || MVK   8, B0
      LDW   *A0, A1
```
No bank conflict:
```
      MVK   0, A0
   || MVK   4, B0
      LDW   *A0, A1
   || LDW   *B0, B1
```